Why did my RAG answer miss the exception to the rule?

Somewhere in your chunking step, the rule and its exception landed in two different chunks, and only one of them came back. A chunker that splits on paragraph or token count doesn’t know a sentence three lines down changes everything above it, so it cuts wherever the boundary falls, not where the meaning does. The retriever then scores each chunk against your question independently. The chunk with the general rule matches well; the chunk with the narrow exception, if it doesn’t share the query’s vocabulary, may not surface at all.

The model isn’t hallucinating here. It’s being faithful to exactly the context it was handed, which is what makes this failure invisible to any check that only grades the answer against its own context. One writeup of a compliance RAG pipeline traced this exact failure to a chunk that ended mid-sentence, right before the exception clause. The fix lives in chunking and retrieval, not in the model.